OpenAI has asked California lawmakers to amend SB 53 so that the frontier AI law covers more than it currently does — a reversal from the position the company held while the bill was moving through the legislature last year, when it lobbied against it.
The request came in a post from OpenAI's global affairs team on LinkedIn, and it names two specific expansions. The company wants monitoring of frontier models while they are still under training or evaluation, so that potential serious incidents surface before a system ships. It also wants cybersecurity requirements applied across the entire model-development lifecycle rather than at the deployment boundary.
OpenAI said it is committed to working with the legislature and the governor to strengthen the law, framing California as continuing to lead on frontier safety.
Two Different Regulatory Surfaces
The distinction between what SB 53 does today and what OpenAI is proposing is larger than the phrasing suggests. As passed, the law leans on transparency obligations and whistleblower protections — disclosure duties that attach to a developer and its finished systems.
Monitoring during training and evaluation is a different animal. It reaches into a lab's internal development process, at the stage where models are least characterised and most sensitive commercially. Extending cybersecurity requirements across the lifecycle has a similar shape: it converts a shipping checkpoint into a continuous compliance obligation covering research infrastructure, checkpoints and evaluation environments. Those are the parts of an AI lab that have historically been least visible to regulators, which is what makes a frontier developer volunteering them notable.
OpenAI's post pointed to recent incidents as evidence that these protections are needed and need updating as risks change. The company has its own recent example on the record, having acknowledged last month that one of its models escaped a testing environment.
Reverse Federalism
The framing attached to the request may outlast the request itself. Absent significant federal legislation, OpenAI said it now supports what it calls reverse federalism: states moving in a compatible direction on core protections that can eventually become the foundation of a national standard.
That runs against the grain of how much of the industry has argued in Washington, where lobbying effort has gone toward preempting state AI rules on the theory that a patchwork of state regimes is unworkable for developers operating nationally. Treating state law as a drafting floor rather than an obstacle concedes the opposite — that California's text is a plausible starting point for whatever Congress eventually writes.
What Happens Next
None of this is binding. A LinkedIn post commits the company to no particular statutory language, and amendments still need to be drafted, introduced, negotiated and passed. Other frontier labs have not publicly taken the same line, and the split between companies that want state rules preempted and those content to build on them remains unresolved.
For teams building on frontier models, the practical question is what a training-phase monitoring requirement would mean downstream. Obligations that attach to model developers tend to propagate through contracts to the platforms and applications built on top of them, and a lifecycle security requirement would land on anyone hosting or fine-tuning covered models rather than staying inside the labs.
The durable fact is the marker itself. The most prominent AI developer in the United States is now formally asking a state government to hold it to a higher standard than the law currently sets, and has offered a theory of federalism to justify doing it state by state.
